American Association for Aerosol Research - Abstract Submission

AAAR 39th Annual Conference
October 18 - October 22, 2021

Virtual Conference

Abstract View


Improving Quantitative Analysis of Spark-Induced Breakdown Spectroscopy: Multivariate Calibration of Toxic Metal Particles Using Machine Learning

HANYANG LI, Leonardo Mazzei, Christopher Wallis, Anthony S. Wexler, University of California, Davis

     Abstract Number: 44
     Working Group: Aerosol Standards

Abstract
We have recently developed a low-cost spark-induced breakdown spectroscopy (SIBS) instrument for in-situ analysis of heavy metal particles nebulized from aqueous solutions. In this work, we investigated the application of machine learning methods to improve the quantitative analysis of elemental mass concentrations measured by this instrument.

Compared to the classical univariate calibration approach used in spectral analysis, multivariate calibration approaches have been found to be beneficial to eliminate matrix effects and enhance analytical sensitivity. In this context, we applied the machine learning methods of least absolute shrinkage and selection operator (LASSO), partial least squares (PLS) regression, principal component regression (PCR), and support vector regression (SVR) to develop multivariate calibration models for 13 types of toxic metals, some of which are included on the US EPA hazardous air pollutants (HAPS) list (e.g., Cr, Cu, Mn, Fe, Zn, Co, Al, K, Be, Hg, Cd, Pb, and Ni). The performance of the proposed models was compared to that of univariate calibration models for each analyte, using coefficient of determination (R2) and root mean square error. By computing the limit of detection (LOD) derived by the proposed models, we found that multivariate models tend to have lower LOD than univariate models.

Furthermore, we assessed the applicability of the proposed models for quantifying elemental concentrations in mixtures of toxic metals, serving as independent validation datasets. Among the various models, the mass concentration predicted by the LASSO model shows the best agreement with known concentrations of each analyte. In contrast, SVR performs relatively poorly. Ultimately, the LASSO model developed in this work is a very promising machine learning approach to quantify mass concentration of toxic metals using our recently-developed SIBS instrument.